[llvm] [AMDGPU] Limit promoting allocas that have users with dynamic index above a threshold on number of elements (PR #170327)

[Carl Ritson via llvm-commits]

https://github.com/perlfu commented:

Is the fundamental limit here actually the number of elements in the GEP type, or rather the width of the GEP type in 32b VGPRs?
I guess alignment (or rather misalignment) drives the complexity explosion?

https://github.com/llvm/llvm-project/pull/170327